Anthropic 发布 Claude 3.7 Sonnet

2024 年底到 2025 年初，推理模型成了新的兵家必争之地。OpenAI 的 o1 证明了「让模型多想一会儿」能显著提升复杂任务的表现，但它带来一个尴尬的选择题：用户必须在「反应快但不那么聪明」和「很聪明但等得久」之间二选一，切换还要换模型。这个选择题，被 Anthropic 在 2025 年 2 月用一种巧妙的设计化解了。

Claude 3.7 Sonnet 发布了，它叫自己「混合推理模型」。同一个模型，默认是标准模式，响应迅速、像普通对话一样流畅；需要深度思考时，打开扩展思维模式，模型会自主规划、逐步推理，思考时长还可以配置。快与慢、直觉与深思，不再是两个模型的属性，而是同一个大脑的两种档位。

对开发者的体验提升是立竿见影的。写代码时，简单任务可以秒回，复杂重构可以开足思考马力，不用再为「该用哪个模型」做选择题。Claude 3.7 同时是当时最强的编程模型之一，与 Claude Code 同步推出，在编码场景里如虎添翼。API 里新增的思考预算参数，也成了后来所有推理模型的标配。

混合推理的意义，在于它重新定义了旗舰模型的形态。它宣告：推理能力不该是少数模型的专利，而应该是每个模型的默认能力——要不要用、用多久，交给用户决定。这个设计思路，被后来的所有厂商吸收。

回看 Claude 3.7，它没有制造 o1 那样的「推理时刻」，但它做了一件更实用的事：让推理从「特供」变成「标配」，从「换模型」变成「拨开关」。当一个普通用户随手打开扩展思考，让模型多花几秒想清楚再回答时，他未必知道——这个让「聪明」变得触手可及的设计，是 Anthropic 在 2025 年春天留下的。

From late 2024 into early 2025, reasoning models became the new battleground. OpenAI's o1 proved that "letting the model think longer" dramatically improved complex tasks, but it created an awkward choice: users had to pick between "fast but not so smart" and "very smart but slow," and switching required changing models. Anthropic dissolved that choice in February 2025 with a clever design.

Claude 3.7 Sonnet shipped, calling itself a "hybrid reasoning model." The same model, by default in standard mode, responds quickly and flows like ordinary conversation; when deep thought is needed, switch on extended thinking mode and the model plans autonomously and reasons step by step, with configurable thinking time. Fast and slow, instinct and deliberation, were no longer properties of two models but two gears of one brain.

The developer experience improved immediately. When writing code, simple tasks could be answered instantly and complex refactors could run at full thinking power—no more choosing "which model should I use." Claude 3.7 was also among the strongest coding models of its time, launching alongside Claude Code and excelling in coding scenarios. The thinking-budget parameter added to the API later became standard on every reasoning model.

The significance of hybrid reasoning is that it redefined the shape of the flagship model. It declared that reasoning should not be a privilege of a few models but a default capability of every model—whether to use it and for how long, decided by the user. Every vendor absorbed this design.

Looking back, Claude 3.7 did not create an o1-style "reasoning moment," but it did something more practical: it moved reasoning from "special offer" to "standard equipment," from "change the model" to "flip the switch." When an ordinary user casually turns on extended thinking and lets the model think a few seconds longer before answering, they may not know—this design that made "smart" feel within reach was left by Anthropic in the spring of 2025.